Clustered Multimedia
NOD
:
Popularity-Based Article Prefetching and Placement
Y.J.Kim, T.U.Choi, K.O.Jung, Y.K.Kang, S.H.Park, Ki-Dong Chung
Department of Computer Science, Pusan National University, Korea
Abstract
According to the current profound development of
multimedia and networking technologies, the way people
communicate with, naturally, has evolved from a text-
oriented into a multimedia-oriented. But, the VOD system
technologies can't be easily applied to MNOD system
because of the difference of the basic data and its properties.
This is the reason why we propose the MNOD system.
NOD data composing news articles make a difference
to VOD data in terms of media type and size, life cycle of
articles and frequency of clients' interaction. Because of
NOD data's intrinsic characteristics, NOD article popularity
model may be different from that of VOD videos. Hence, we
analyze statistically the log data of one electronic newspaper
and show the popularity distribution of articles is different
form Zipf
s,
which is a popularity model of VOD data. We
propose a new article popularity model for NOD data, which
we call Multi-Selection Zipf distribution.
Also,
we propose
the article prefetching policy based on the popularity model
and life cycle model of NOD articles for increasing
performance in a
MNOD
system.
And, we consider the data placement on the MNOD
system with prefetch cache. The user requests can be
serviced from a cache
or
a disk. Consequently, in order
to
place a NOD data, we must consider the correlation between
disks and cache according to data popularity.
In
this paper,
we propose the data placement policy that considers both
disk-existence probability and cache-existence probability of
a data using data popularity
1.
Introduction
1.1
Motivation
As a VOD system gets more popularity these days, a
NOD(News On Demand) system supporting multimedia
services, we call a
MNOD
(Multimedia
NOD)
system in this
paper, will be popular as a news service in the near future.
Furthermore, the developments of computer network and
multimedia technology
are
making the multimedia-oriented
news services possible.
Previous works generally regarded VOD and MNOD
systems as an identical category from the viewpoint of using
multimedia data. Multimedia data usually need
high
bandwidth and massive storage space and have real-time
criteria. Therefore, we basically utilize the research results of
a VOD system. Clearly,
a
MNOD system can be categorized
into the same multimedia application as a VOD system, but
it has some intrinsic aspects that a VOD system doesn't
possess[4]. First, the articles of NOD service are made on
and off in a day while the VOD data is made once a half-
month or a month. Second, the more recent article is made,
the more users access the article. The good movies are
preferred to the others for a long time, but NOD data aren't.
The NOD articles, which have the best popularity, don't last
more than three days. Third, the number of articles that are
requested from each user is varied. Owing to
the
long length
of video data, users serviced in VOD system select just one
or two video data. But the length of NOD data is short, as a
result, users
can
select several data at a time. Finally, NOD
data
have the temporal and spatial access locality. The user
requests burst according to the access time and the kind of
articles.
So
the VOD system can't be fully applied to a
MNOD system because of these different characteristics.
In
short, we use the VOD system research results basically, but
will
revise and decorate it.
The previous work showed that the system with
prefetching has the better performance than the system
without prefetching[ 1, 151. But without considering the other
characterizations such as short-term life cycle, the difference
of user's access pattern according to time and etc, we can not
expect to
the
increased performance in the MNOD system.
The system with prefetching only based on data popularity
has the problem, which is data replacement happens too
often. Hence, using the analysis of electronic newspaper log-
files, we propose the popularity model and the life cycle
model of NOD articles. And we will suggest the article
prefetching policy based on time window according as the
models.
And then we consider the data placement on system
with prefetch cache. The user requests can be serviced from
a cache or a disk. Consequently, in order to place a MNOD
data, we must consider the correlation between disks and
cache according to data popularity.
In
this
paper, we will
propose
the
data placement policy that considers both disk-
existence probability and cache-existence probability of a
data using data popularity.
The remainder of
this
paper is organized as follow: In
section
2
the overall architecture of a proposed MNOD
system is presented briefly. Section
3
shows the popularity
model and life cycle model of NOD articles via the statistical
analysis of log files of electronic newspapers, and
in
section
4 we propose the article prefetching policy based on the
popularity and life cycle model of NOD articles. In section
5
194
1-9173/99
$10.00
0
1999
IEEE
the article placement policy in
the
MNOD system with
prefetch cache is proposed and evaluated. Lastly, in section
6
we mention the conclusion and future works.
1.2
Related
work
0
Data caching
and
Prefetching
In
order to establish caching policy for NOD data, the
popularity and life cycle model are important conditions.
[lo] showed that the popularity distribution
of
real video
was a double exponential distribution rather than a Zipfs and
proposed cache policy based
on
the life cycle of videos.
In
[
113, the long-term model of movie popularity was proposed
for a hierarchical VOD system, which was the modified
exponential distribution with three parameters. The
strategies of these papers are not acceptable for
MNOD
system because these analyzed only a VOD data, whose
life
cycle is longer than that of NOD article. [9] made a
psychological analysis
of
access pattern
of
users, based
on
a
model
on
human memory and insisted that the recency of
data was more important criterion in data caching algorithm
than the frequency. But because of various characteristics of
NOD articles, considering only access recency is insufficient
for caching or placement
of
articles.
[4]
proposed a news
article prefetching policy but had not the correct criterion for
the data replacement.
In
this
paper, we will show the measure of short-term
and long-term
of
article popularity. And we will propose
article prefetching policy considering that measures.
0
Data placement
In
Multimedia system, data placement policies have
been studied
so
as to increase the success ratio
of
user
requests. Therefore, most studies were continued
to
get the
high
disk
bandwidth. Data Striping method is the
representative placement policy. The famous striping
methods are round-robin, random[3], staggered[7],
permutation striping[8] and etc. And then, the policies to
prevent disk bottleneck, such as dynamic replica policy[5]
and data placement policy based
on
popularity were
proposed.
But these policies considered only
disk
conditions
based
on
data popularity. But
in
these days, since almost
all
systems have data cache, the user requests are serviced from
a cache or a
disk.
Consequently,
in
order to place data, we
must
consider
the
correlation
between
disks
and
cache
according
as
data popularity.
2.
MNOD
system
overview
Though
a NOD(News
On
Demand) system has many
differences from the well-known VOD(Video On Demand)
system, a NOD system is
still,
in
many aspects, similar to a
VOD system in terms of using multimedia data. Hence, we
can draw out the structure of a MNOD system using that of a
VOD system. And the scalability of an MNOD system has to
be investigated to support hundreds
of
thousands of
simultaneous users.
As
shown
in
Figurel, we suggest the
two-layered clustered MNOD server system by considering
the scalability. The suggested system structure contains two
main modules: a
control node
and a
storage node.
Many
studies
on
the VOD system architectures, theoretically,
assume that there are
two
logical parts in the server[l,
2,
151.
Roughly speaking, the control node deals with the admission
control for the storage node and the storage node schedules
the admitted requests from
the
control node and treats
storing and retrieving data,
The control node is entrusted with the control
part
of
tfie whole
MNOD
server. It controls the permission into
entering the storage node. If a NOD saeam arrives at the
control node, the admission manager checks if the requested
stream may be admitted
or
not and whenever a request is
permitted, it is passed from the control node to
the
storage
node. The storage node is responsible for both storing and
retrieving NOD data. Now that the storage node only
schedules the permitted streams from the control node,
no
(x>
CI??Il
e..
cz3
Client Client Ciienr
Reques!
1
.Reauest
-NO3
.1
Fos:
Network
,
1
Control
Node
3
..
Control
Node
\
Admitted
Request
1
Admitied
Request
1
Storage
Nods
Figure
1. Proposed clustered
MNOD
system
special admission control
is
required.
3.
Article popularity model
NOD data composing news
articles
make a difference
to
video data of
VOD
in
terms
of
media
type
and size, life
cycle
of
article and frequent clients’ interaction etc. And,
NOD data may be created
on
and
off
at any times, and it is
popular and recent articles that most
of
the clients access
mainly. Because
of
these characteristics, NOD article
195
popularity model may be different from that of VOD movie.
In
this
section we
give
a statistical analysis of log data
of electronic newspaper running currently, and show that the
popularity distribution of article is different from Zipfs,
which is a popularity model of VOD.
In
particular, we
propose a new article popularity model, which we call
Multi-Selection Zipf distribution,
and the algorithm
generating its probability distribution.
This
article popularity model is a short-term popularity
model that denotes popularity distribution among articles in
an instant time. While the short-term popularity model
notifies what articles are prefetched, long-term popularity
models, denoting the life cycle of articles, notify when the
articles are prefetched. In addition, these models can be used
for caching and placement
of
articles. Thus,
in
this
section
we also present a long-term popularity model of NOD
articles.
3.1
Multi-Selection Zipf distribution
0‘06
0.05
0.04
0.03
0.02
0.01
0
1 10 19 28
37
46
55
€4
73
82
91 100109118127136 145
(a) Average access popularity
0.16
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
-NOD
data
--zlPf
.......
Exponential
-mtmo
mr-mo
mhV)z-mr-V)
~oGoaacEamg_c~zg
(b)
Comparison with other distributions
Figure
2. The real access distribution of NOD article
The typical popularity model
of
VOD
data
is
Zipf
distribution; it denotes access probability of the i-th popular
movie when
the
movies are sorted
in
the order of decreasing
popularity. Then, do the popularity of NOD data follow a
Zipf distribution? We analyzed statistically the log files of
electronic newspapers running currently and found that news
article’s popularity model is different from Zipf
s.
In Figure2, (a) shows
the
mean access popularity of
news articles for a month, resulted from making a statistical
analysis of log data of June
1998.
(b)
shows the comparison
of NOD article popularity with Zipfs and Exponential’s;
The upper portion
of
NOD article distribution decrease less
drastically than Zipfs but more fast than Exponential’s.
When a user connects with a NOD server, the user requests
several articles rather than only one article.
This
imply that
popularity does
not
concentrate on one article but spread
over several articles. Consequently the curve of popularity
distribution of NOD articles is less steep than Zipfs. For
comparison, we use a pure Zipf distribution without skew
and
an
exponential distribution with
/z
=0.01.
A
user does not watch only one article but several
articles while he
is
connected with a NOD server. If a user
selects
k
articles on average in NOD server with
N
articles,
the possible number of
k
article selections
is
as follow:
M=[T)=
N!
k!(N
-k)!
In other words it may be thought that he selects a group
composed of
k
articles in
M
article groups. Suppose that the
popularity of these Marticle groups follow a Zipf
distribution,
the
probability
on
which each
article
is selected
is the normalized sum of probabilities of groups containing
the article. The definition of Multi-Selection Zipf
distribution and the algorithm generating the probability
distribution are as follows:
Definition:
Suppose that a user selects
k
articles
on
average
in NOD server with
N
articles and the popularity of the
M
article groups follows a Zipf distribution, we define
article access probability as the normalized sum of
probabilities of article groups containing each article. We
call the distribution of article access probability
Multi-
Selection Zipf distribution
Input:
k
(the number of selections),
Algorithm:
N
(the number
of
article)
Generate
M
article groups;
Decide popularity of the groups and
sort
in
Popularity order;
Calculate the popularities of groups using
Zipf distribution;
while
(all
articles)
{
adding the probability of groups containing
the article;
normalize such that the sum of calculated
probabilities is
1;
1
algorithm
Figure
3.
Multi-Selection Zipf probability generation
Figure4 compares Multi-Selection Zipf distribution
with Zipfs; (a) shows that the curve of Multi-Selection
Zipfs is less steep than Zipfs.
In
(b), as the number of
196
0.15
O'?
i
(a)Comparison of k=5
with
Zipf
0.14
I
0.12
0.1
0.08
0.06
0.04
0.02
0
Figure
4.
Multi-selection Zipf distribution and Zipf
s
0.12
-
0.2:
hour
(a) Access Popularity as article life cycle
(b)
Life cycle of
two
articles
Figure
6.
Article life cycle
Figure
5.
Multi-selection Zipf distribution
&
the real
distribution
of
NOD
article
selections
is
3
and
5,
It shows that the distribution
with
5
prefer new articles. Long-term popularity model, called as
selections decrease slower than
3
selections. That implies article life cycle, reflects the changes
of
the popularity of an
that the more selections are, the less steep the curve
is.
article.
In
Figure6, (a) shows a long-term popularity
Figure5 shows the comparison of the real popularity resulted distribution as a result of statistical analysis of log files; it
from
log
files with Multi-Selection Zipf
s
with
5
selections. has left skew, which shows that users prefer new articles to
the old
ones.
3.2
Long-term
popularity
model
Long-term popularity model can be used for deciding
the time
to
prefetch
an
article into the cache
to
decrease
the
retrieval overhead. In Figure6
@),
the popularity of article
A
decrease and article
B
increase, and the best point that the
The popularity of articles changes as time passes
because new articles are generated at any time and users
197
article
A
should be replaced with article B in the cache
is
around the hour 10, when the curves of two articles meet.
4.
Article prefetching policy
In previously suggested clustered MNOD server,
clients request articles from CSVR(Contro1 Server), and
CSVR fetches articles from SSVR(Storage Server) and
sends them to clients. If popular articles are prefetched in
CSVR, the service delay can be reduced and network traffics
can be decreased by eliminating the time to retrieve articles
from SSVR[lS].
In prefetching popular articles, old articles must be
replaced with new popular ones when new articles are
generated and the popularities of old articles decrease.
So,
we need to consider two popularity models mentioned in the
previous section for better performance in article
prefetching. In this section we present the new prefetching
measure combined two popularity models and suggest the
article prefetching algorithm employing the measure.
4.1
The measure of prefetching
We express
the
measure of short-term article popularity
in
terms
of the access frequency during some time intervals
and
the
measure
of
long-term article popularity in terms of
the increased or decreased amount of frequency during some
time interval. For example, the article whose access
frequency increase rapidly may be a new article while the
article whose access frequency decrease may be less old one
and
the
article that access frequency hardly decrease may be
old one.
Note that we use time window whose interval is several
ten minutes and calculatefi, the access frequency, by
cumulating references
to
the article during any window
i.
By taking a weighted average of fluctuations in all the
windows, we can calculate
ai,
the expected variation
amount of access frequency in next window as follows:
In Eql,
sf
,
denoting a smoothing factor, is the weight
that indicate the importance of frequency variations of
current window; The larger
Sf
is,
the heavier the weight of
the current window’s frequency
is.
Consequently we propose
a new measure (Life-cycle
Based
Frequency) by
combining
fi
with
ai
asfollows:
LBFi
=
fi+
ai
m.21
4.2
Prefetching scheme
CSVR maintains
logs
of
reference to each article
during each window. From the logs, access frequency Aof
each article in the window can be calculated. Compute
aiusing [Eq.l] and
DF,
of
each article using pq.21.
Because the article that
has
the largest
LBF,
is
expected to
be accessed most frequently in next window, the articles that
have large
LBe
have to be prefetched. Once articles
are
prefetched in CSVR, the articles can’t be replaced with
others during current window. Figure7 shows the prefetching
algorithm.
while (1)
{
if (a new window)
{
while (for
all
articles
in
window)
{
“i
=
Sf
(
fi
-
A.,)
+
(1
-
sf
)“&I
DF,
=J.+q
1
sort the articles in
LBF,
order;
prefetch the articles that have large
LBc
until
cache memory is
full;
:
Figure
7.
Article prefetching algorithm
43
The effects
of
prefetching
In the proposed clustered
MNOD
server, the
prefetching in CSVR promotes the efficiency of the server
system because of decreasing the overhead of retrievals from
disks in
SSVR.
We employ the
miss
rate and the concurrent
users to show the effects of prefetching. For simulation, we
used the log files of Pusanilbo’s electronic newspaper in
Korea, and the parameters expressed in Table1
.
Parameter Values
Block size
Figure8 shows
the
relations between
the
amount of
prefetching and the
miss
rate; as the amount of prefetching
increases, the
miss
rate decreases. Figure9
also
shows that as
the prefetching amount increases, the number of concurrent
users by LBF scheme and ones only including prefetching
overheads. Because prefetching
an
article is equal to
processing a request, the overheads of article prefetching
are
calculated
by
multiplying
the
number of prefetched
articles
and
the overheads that a request requires. Thus the real
effect of article prefetching
is
expressed
as
the difference
between the number of concurrent users by
LBF
scheme and
198
the number of concurrent users only including prefetching
0.8
r
::;
1
.
0.5
1
0.4
1
0.3
1
0.2
t
t.
mss
rate
i
0.1
1
0
100M 200M 300M 400M
500M
Figure
8.
Miss
rate according to prefetching size
1
40
r
20
I
+-
Cuncurrent
Users
+
C.
U
including
the
overhead
01
OM l00M 200M 300M 400M 500M
Figure 9. Concurrent users according to prefetching size
overhead.
As
a result, prefetching articles in
CSVR
increases
the performance of MNOD server system by 1.5 times.
5.
Article Placement Policy
5.1
previous
Data Placement
We intended
to
increase user service throughput by
arranging replica of popular data
on
disks.
This
method,
however, had a few problems.
NOD article is not as big
as
VOD data
in
size.
In
MNOD system, relatively large amount of data exists
in
the
cache, and the probability that hot data
is
in
the cache
is
higher than VOD system. Naturally, disk access probability
will
be lower relatively. And, the number
of
NOD articles
is
very much than VOD data, and by reason of storage
overhead the number of articles being replicated needs to
be
restricted
to
articles
with
high
popularity.
Thus,
the data that has lower popularity but more
growing probability of disk access
will
be suffered from
long service delay because of small number of replica due
to
low popularity.
5.2
New Scenario
of
Data Placement
The new scenario we suggest for data placement
in
the
MNOD system is to manage various
numbers
of replicas
according to data popularity.
The article data with higher popularity will be cached
by some caching method based
on
data popularity and have
lower probability
of
disk access. The article data that has
lower popularity but more growing probability
of
disk access
needs higher disk
YO
bandwidth.
So,
more many
of
replicas
need to be assigned for article data with lower popularity.
By reason
of
this,
we emphasize the article data that
has relatively lower popularity than some hot data.
5.3 Article Placement Algorithm
53.1 Determining the number
of
article
replica
Article data is arranged in decreasing order according
to
its
popularity and has
an
identity number
of
its
own
that is
able to be discerned from other ones. We
will
bestow the
replication level,
T,
,
on
article data
ai
in order to determine
the number
of
replica.
The replication level
T,
is
the inverse of difference
between mean article popularity and popularity value
of
article data.
This
algorithm also takes advantage of standard
derivation for assigning largest value of
Ti
to the data that
has to be repeated
many
times. And, by that result, the
average
is
shifted.
We can express
T,
as
[Eq.3]
by integrating the
expressed mentions above
1
-15
I
c
I
15
[Eq.
31
IW)
+
X)
-
P(X,>
1
’
where
x2
:
identity of article data
p(x,
)
:
popularity of
Xi
X
E(X)
a(
X)
:
random variable of
p(x,)
:
expected value (mean) of
X
:
standard deviation of
X
533
Article placement
on
disk
Article data consists
of
several data blocks that
are
units
of placement. We place article data
on
the disk using
round-robin method and the disk where the first block of
news data is placed
is
selected randomly.
If replicas of article data exist, we will place them
using the same scheme, but we must utilize
a
random seed
for
seeking
a
start
disk
that is not the same disk storing
original article data,
so
as
to secure higher disk
I/O
bandwidth from
disk
access collision. Figure10 shows the
article placement algorithm proposed
in
this
paper and the
199
Thus, maintaining the fully replicated popularity level to the
higher level 10%-20% makes the best performance.
0.3
I
1
.?.’5
E--i
50.;
-,*
,:
.I
..’
0.05
0
0
i5
30
45
60
90
Amount
of
users
1
Figure 12. Service failure to amount of users
I
~
O3
I
025
I
I
I
I
It
O2
E
1
i
d
g
0’5
01
I
0
05
0
0%-10%
(00)0%-10% (00)
10%
10% 20%-30%
10%-30%
-20%(00)
-2O%(Eo)
190)
(EO)
Fully replicated popularity
level
Figure 13. Service failure to every disk
in
caching
wqtern
Figure13
present service failure rates of each disk when
the number of concurrent users is 60 and 90. The results
show that the proposed article placement policy controls the
load balance of every disk as well.
6.
Conclusion
In
this
paper we propose popularity-based article
prefetching and placement policy for NOD article data.
Because NOD data composing news articles make a
difference to video data of VOD
in
terms of media
type
and
size, fife cycle
of
articles and frequency of clients’
interaction, a new NOD article popularity model
is
needed to
explain user access patterns.
As
a result, we show that
the
popularity distribution
of
articles
is
different from Zipfs,
which
is
a popularity model of VOD data.
In
particular, we
propose a new article popularity model, which we call Multi-
Selection Zipf distribution. The proposed model
is
the
short-
term
popularity model that denotes popularity distribution
among articles in an instant time. If the article prefetching
algorithm may be considered, while the short-term
popularity model notifies what articles are prefetched, long-
term popularity model, denoting the life cycle of articles,
announces when the articles are prefetched. Thus, we
propose the article prefetching policy based
on
the
popularity and life cycle model of NOD articles for
increasing performance
in
a
MNOD
system.
And, we propose the data placement policy for NOD
data that considers both article popularity and disk access
probability.
In
this
policy, we replicate data that have not
best popularity, but high disk access probability because the
probability that the best popular data is in cache memory is
high and relatively it’s disk access probability is IOW. And,
original data and replica are placed to minimum loaded disks
to achieve load balancing among disks. The proposed article
placement policy decreases the service failure rate and
increases the service throughput by way of load balancing
among
disks.
Our future work is to develop a mathematical model of
Multi-Selection Zipf distribution and to investigate article
placement and prefetching algorithms in clustered
architecture environment.
Reference
[l]
Louis
P.
Slothouber,
“A
Model of Web Server
Performance“,
http://vorlon.biap.com/webperformance/
modelpaper.html, 1996.
[2]
Odysseas I. Pentakalos, Daniel
A.
Menas&,
“An
Approximate Performance Model of a Unih-ee Mass
Storage System”, 14” IEEE Symposium
on
Mass Storage
Systems, pp. 210-224,
Sep.1995.
[3]
J.
Alemany,
J.
S.
Thathacher, “Random Striping for
News on Demand Servers”, Technical Report UW-CSE-
97-02-02, Univ.
Washington,
1997.
[4]
Y.
U. Park,
G
H.
Back,
W.
I.
Seo,
Y.J.
Kim,
K. D.
Chung,
“New Buffer policy for NOD data”, Conference
on Korea Broadcast Engineering, Korea, 1997.
[5]
A. Dan, M. Kienzle, D. Sitaram,
“A
dynamic policy of
segment replication for load balancing in video-on-
demand servers”, Multimedia System 3:93-103, 1995.
[6] T. D. C. Little,
D.
Venkatesh,
“
Populary
-
based
assignment of movies to storage devices
in
video-on-
demand
system”,
Multimedia
System
2:280-287,
1995.
[7]
S.
Berson, R.
Muntz,
S.
Ghandehaxizadeh,
X.
Ju,
“Staggered Striping
in
Multimedia Information
Systems”,
SIGMOD Conference, pp. 79-90,1994.
[8]R. Flynn,
W,
Tetzlaff, “Disk Striping and Block
Replication
Algorithm
For Video File Servers”,
IEEE
3d
International Conference
on
Multimedia Computing
Systems, 1996.
[9] J.
E.
Pitkow,
M.
M. Recker,
“A
Simple Yet Robust
Caching Algorithm Based
on
Dynamic Access Patterns”,
20
1
In the Proceeding of the Second International
WWW
Conference, 1997.
[lo]
S.
A. Barnett,
G
J.
Anido, H. W. Beadle, “Caching
Policies in a Distributed Video-on-Demand System”, In
the Proceeding
of
Australian Telecommunication
Networks and Applications Conference, 1995.
[ll] C. Griwodz,
M.
Bar,
L.
C. Wolf, “Long-term Movie
Popularity Models in Video-on-Demand Systems or
The Life of an on-Demand Movie”, The
fifth
ACM
International Multimedia Conference, 1997.
[12]
I.
Tatarinov,
V.
Soloviev, A. Rousskov, “Static Caching
in Web Servers”, submitted to
the
6*
international
Conference on Computer Communications and
Networks.
[13] Lee
D.,
Choi
J.,
J.
Kim,
S.
H.
Noh,
S.
L.
Min,
Y.
Cho,
C.
S.
Kim,
“LRFU Replacement Policy: A spectrum
of
block replacement policies”, In Seoul National
University Technical Report SNU-CE-AN-96-004,
Korea, 1996.
[14] Renu Tewari, Harrick M.
Vi,
Asit Dan
&
Dinkar
Sitaram,
“Resource Based Caching for Web Servers”,
Proceedings of SPIEIACM Conference on Multimedia
Computing and Network, 1998.
[15]
Renu Tewari, Rrajat Mukhejee, Daniel
M.
Dias,
Hanick M. Vin, “Real-Time Issues for Clustered
Multimedia Servers”, IBM Research Center, Technical
Report, 1995.
202